I want to be clear that I did not go through these stages. I saw them coming, rationally assessed the trajectory of the technology, and made a calm, evidence-based decision to embrace AI tooling early. This is my official position and I am sticking to it.
My Slack history from early 2023 is not admissible evidence.
Stage One: Denial (Average Duration: Two Weeks)
The denial stage has a very specific texture. The engineer in denial is not ignorant of AI — they are informed, articulate, and specifically dismissive. They have read the papers. They have opinions about hallucination rates. They will tell you, at length, that large language models are “just statistical pattern matching” as if this somehow makes the output less useful when it’s faster and more accurate than what they’d produce manually.
The tell-tale sign of Stage One is the phrase “it’s just autocomplete.” This is technically partially true in the same way that a Formula One car is “just a vehicle.” True at a mechanical level. Entirely missing the point in every way that matters.
The denial stage ends, almost universally, the first time the engineer asks the model something they expected it to get wrong — and it doesn’t.
Stage Two: Anger (Average Duration: One Week)
Having accepted that the model can do some things, the Stage Two engineer devotes considerable energy to cataloguing what it cannot do. This is actually useful research, conducted with entirely the wrong intent.
The Stage Two engineer has found the hallucination. They have found the bad SQL. They have found the confident wrong answer and they are brandishing it at everyone in the vicinity like a golden ticket. Look. It got this wrong. Therefore the technology is fundamentally limited. Therefore I am safe. QED.
The problem with this logic is that “sometimes wrong” is a description that applies to every tool that has ever existed, including the Stage Two engineer themselves. The relevant question is not “does it make mistakes” but “does it produce net value compared to the alternative.” The Stage Two engineer is not ready for this question.
The engineers who spend the longest in Stage Two are usually the ones who are best at the specific tasks the model does adequately. There is a real grief here that deserves acknowledgment: if you’ve spent years developing expertise at something and a model can now do it adequately in thirty seconds, that is genuinely disorienting. The anger is understandable. It’s just not a strategy.
Stage Three: Bargaining (Average Duration: Three Days)
Bargaining is the shortest stage and the funniest to observe from the outside. The bargaining engineer has accepted that the model is useful but is negotiating the terms of coexistence. They will use it, but only for documentation. Only for boilerplate. Only for the boring stuff — the real work is still theirs.
This lasts approximately three days because the boring stuff turns out to be 70% of the job. Once automation removes it, the calculus changes immediately.
Stage Four: Depression (Average Duration: One Existential Week)
This one is real and worth taking seriously.
The depression stage is not about the technology. It is about identity. Engineers — particularly experienced ones — have often built a significant portion of their professional self-image around their ability to produce code. Fast, clean, expert code. When a model starts producing code of comparable quality in a fraction of the time, the question that surfaces is uncomfortable: what is my value now?
The answer is: judgment. Context. The ability to ask the right question, catch the wrong assumption, architect the thing that needs to be built rather than implement the thing that was asked for. These are not skills the model has. They are skills that have been undervalued because the implementation work obscured them.
The depression stage ends when this becomes clear. For some engineers it takes a week. For some it takes a jarring production incident where the model’s output — accepted without review — did something technically correct and domain-wrong. Nothing clarifies the value of human judgment like watching what happens without it.
Stage Five: Acceptance (Duration: Indefinitely)
The engineers who reach acceptance stop talking about AI as a topic and start talking about what they built. The tooling is infrastructure now. It is not discussed any more than a text editor is discussed. It is used.
Their output is different. Their PRs are more complete. Their documentation exists. Their test coverage is higher because writing tests is no longer a friction-heavy task that gets deprioritised. They spend their time on architecture, requirements, and review — the parts that were always most valuable and are now, finally, getting the time they deserve.
The Stage Five engineer is not more productive because they have a better tool. They are more productive because they have stopped spending skilled human hours on tasks that never required skilled human hours in the first place.
I did not go through the stages. But if I had — and purely hypothetically, if I had — Stage Four would have been the important one. The question of what actually constitutes expertise in a world where pattern execution is automated is worth sitting with.
The answer, once you find it, is clarifying. And then you get back to work. There’s a lot of it. Ship faster.